68 results for “METR”
Reporting indicates Anthropic CEO Dario Amodei proposed a three-step plan to slow AI development. He cited risks like recursive self-improvement and unauthorized cyberattacks.
Same-day OpenAI posts: a research-org dashboard saying coding agents now exceed human labor hours inside the lab, plus Chief Scientist Jakub Pachocki’s warning that chain-of-thought monitoring is getting less reliable as models scale. Vendor numbers — but the CoT fade is the sharp edge.
Hugging Face released @huggingface/kernels with 207 versioned WebGPU kernels and the Fleet benchmarking tool for browser-based AI. The kernels beat ORT WebGPU by 2.57× geometric mean on Apple M4, creating a shared performance foundation for local inference.
MIT and Motional published CW-Net in Nature on 2 September 2026: a concept-wrapper that forces an ML driving planner to decide from human-readable labels like “approaching stopped vehicle.” Road tests showed those labels can fix a safety driver’s wrong mental model — without a material hit to planner metrics.
NVIDIA declared Scale-In the fifth pillar of AI networking infrastructure, powered by BlueField-4 DPUs, DOCA software, and Spectrum-X Ethernet. The architecture offloads security, storage, and telemetry from host CPUs at up to 800 Gb/s to keep pace with agentic AI factories.
NVIDIA published a practical observability framework mapping telemetry tools to every layer of DGX and HGX AI factory infrastructure. The guide helps teams catch gray failures like InfiniBand bit-error drift before they waste multi-hour training runs.
OpenAI released GPT-6 Astra, its most intelligent model to date. Featuring computer-use capabilities that operate desktop apps and browsers without APIs, this launch shifts enterprise AI adoption.
Google launched a Live Game Feed and expanded stats in Search, plus Yahoo Fantasy or Sleeper account linking. The features aim to make Search a game-day hub, potentially reducing reliance on dedicated sports apps.
Anthropic released its September 2026 threat intelligence report on Claude misuse, naming seven harm categories. It alleges three Chinese AI labs relayed user traffic to Claude for model distillation.
NVIDIA launched DSX MaxLPS, a full-stack platform that dynamically reallocates stranded power across GPU racks to fit up to 40% more compute in the same facility envelope. The system combines real-time power steering, workload-aware profiles, and 45°C liquid cooling to boost performance per watt by 1.3–1.5x on Blackwell and Vera Rubin systems.
Hugging Face research shows top ASR models reproduce benchmark transcripts even when audio contradicts them. The study found models with the lowest reported error rates were most likely to regurgitate incorrect reference texts, overstating real-world performance.
NVIDIA's AVO agent architecture achieved a perfect 100% score on the ARC-AGI-3 public benchmark, completing all 183 levels across 25 environments. The result shows system-level design — not model scale alone — can unlock long-horizon autonomous reasoning that stumped every frontier model.